Differentiable molecular simulation can learn all the parameters in a coarse-grained force field for proteins

Area of Science:

• Computational chemistry
• Molecular dynamics
• Biophysics

Background:

• Optimizing molecular simulation force fields is complex and time-consuming.
• Tuning multiple parameters simultaneously presents a significant challenge.

Purpose of the Study:

• To demonstrate the application of automatic differentiation for optimizing molecular simulation force fields.
• To develop a coarse-grained force field for proteins using differentiable methods.

Main Methods:

• Utilized automatic differentiation to obtain gradients of a loss function with respect to force field parameters.
• Performed training simulations to learn parameters for a coarse-grained protein model.
• Validated the learned potential against chemical knowledge, PDB data, and protein folding/dynamics simulations.

Main Results:

• The differentiable molecular simulation approach successfully parameterized a coarse-grained protein force field.
• The learned potential accurately reproduces native protein structures and dynamics.
• Demonstrated utility in protein design and model scoring applications.

Conclusions:

• Automatic differentiation offers an efficient and interpretable method for optimizing molecular simulation force fields.
• This approach integrates deep learning advancements with established simulation techniques.
• The developed methods and potentials are applicable to various protein modeling tasks.